Study On Extraction Of Broad-levaed Forest Information Based On Medium And High Spatial Resolution Remote Sensing Image

As the world's largest terrestrial ecosystem, forest is the necessary foundation of human survival and development, which not only provides rich lumber and Lin by-products, but also plays an important role in climate regulation, water conservation and environment protection. Therefore, for improving decision--making level of forestry development and promoting forestry, social economy and even the global environmental sustainable development, there is a very important significance to grasp the status and changes of forest resources by forest resource inventory. Compared with the traditional survey methods, remote sensing intelligent interpretation based on remote sensing image is widely used in forest resources monitoring, for its microscopy, comprehension, short cycle, repeatability, low cost and other characteristics.In recent years, numerous classification methods were proposed in the literature and utilized in many fields successfully, in which, classification combining spectral characteristics and multivariable remote sensing data has particularly become study focus. In this paper, band characteristics were analyzed to throughly understand ALOS data by single-band and multi-band statistical methods, and spectral characteristics of different vegetation types were computed by adopting Normalized Difference Vegetation Index (NDVI), Principal Component Analysis (PCA) and Optimum Index Factor (OIF). The ideal decision tree classification algorithm was built to study the vegetation classification of Pingnan County in Guangxi Zhuang Autonomous Region, with association of field survey data and comparison of maximum likelihood method. The method was applied to TM and SPOT5 after adjustment, aiming at providing references for rapid information extraction from high resolution remote sensing data. Paper was supported by "high-resolution earth observation system of major projects (E0305/111202)" and forestry research & special public service sectors "stand structure and growth simulation studies (201104028)". The main research conclusions are as follows:(1) With reference to correlation degree and information statistics, correlations between band4 and other bands are lower, indicating that band4 has greater independence and should be chosen in further processing phases, followed by the bands 3,2,1.(2) OIF analysis shows that, the best RGB bands combination is bandl, band3 and band5 (NDVI) in theory, but considering the information amount, object discrimination and color richness, band 4,3,2 combination performs better.(3) Through analyzing four original bands of ALOS, it is found that spectral responses of various vegetations in some wavelengths were similar thus difficult to distinguish. But considering NDVI and PCA, the distinction between different vegetations was significantly enhanced, which provide references for band combination of broad-leaved forest extraction.(4) For ALOS data, the overall accuracy of decision tree classification reached 89.94% with kappa coefficient 0.8787, increasing significantly compared with maximum likelihood classification, for which broad-leaved drawing precision was 88.14%, user precision was 88.89%. According to vegetation spectral features in ALOS, it is suggested that using decision tree classification algorithm to select right threshold extracting vegetation type is feasible.(5) The accuracy of maximum likelihood algorithm based on ALOS reaches 83.90% for broad-leaved forest classification, but when it comes to bamboo, shrub, farmland etc., the error increases. Though still misclassification and missed observations existed, decision tree classification can overcome the deficiencies of mixed classification to a large extent, subsequently improve the accuracy of broad-leaved forest extraction.(6) Confusion matrix analysis shows that, the quality of decision tree classification based on TM and SPOT5 was ideal, the overall accuracy was 86.05% and 90.61% respectively. The results suggest that the method can classify and discriminate vegetation effectively, and have certain universality for vegetation classification with different data source, therefore, provide theoretical foundation and effective means for extracting vegetation automatically.Paper was supported by stand structure and growth simulation studies (201104028).